JAVATEL — IP house across 7 domains

From video, language & AI to device, quantum, biochem & autonomous.

Founded 1997. Patent-anchored, math-backed foundation tech across 7 domains. Video and language are our two main pillars, with AI inference control, device integration, quantum resistance, biochem and autonomous-systems control expanding from the same theoretical base.

A thesis, with working proof ― AI

Sovereignty Without Ownership

A device that seals the atypical judgments born in conversation into deterministic executables, on the spot.

AI sovereignty is not owning the giant model. Borrow the teacher. But as long as you hold the clamp that validates the output — canonicalization, deterministic routing, bit-exact judgment, audit — sovereignty never leaves your hands. And the clamp is light (tiny, WASM), so holding it takes no vast capital.

Routine (standard/repeated, e.g. an expense cap) up front; atypical (sudden/person-dependent, e.g. a year-end exception) added on the spot in plain language. Both are clamped bit-exact and recorded.

The hard parts (bit-exact, lisp-jit) aren't hidden; the surface is a form everyone knows. Put values in the table, say the rule in natural language, and the clamped result comes out with an audit trail. All in-browser, zero upload.

The thesis & full picture →Clamp demo (spreadsheet / templates) →Fold demo →

LANGUAGE · Highlight

Conventional wisdom

No. BLACKBOX — bring it on.

COBOL migration does not need a “semantic understanding” step. Treat the source as the spec, deterministically transcribe it bit-exact. That’s all.

Slime-Boy: Why don’t you need to understand the meaning?! 💢

“Meaning” depends on human perception — there is no mathematical rigor in it. The same code reads differently to different people, so an “understand-it-then-rewrite” migration injects that wobble.

The target is bit-code written as 0101… — that is structure. Structure is unambiguous and uniquely determined.

SlimeNENC projects (π) the structure of the original COBOL onto Slot IR and transcribes it structure-preserving into the target (Java, etc.). No “understand the meaning” step — so it can guarantee mathematical rigor (bit-exact).

Not only legacy — modern languages too. The same structural translation, now Python → Rust, bit-exact (SHA-256 identical): SlimePython →

LANGUAGE category → LANGUAGE services → Latest news →

★ LANGUAGE impact ★ DEVICE I/O cross-listed

Once the migration is done, the next problem is storage.

Even after a BLACKBOX bit-exact transcription succeeds, if the destination is a general-purpose DB like PostgreSQL, financial-core overnight batches often crush their window. SlimeTree-VSAM provides VSAM-compatible (KSDS / ESDS / RRDS) native storage as a single Rust binary — 480× faster than PostgreSQL in same-host bench, with a SHA-256 audit chain built in so tamper detection works even in air-gap environments. A simple-record-system variant, member of the Slime storage family.

SlimeTree-VSAM details → Primary materials / white paper → LANGUAGE category →

VIDEO · Highlight

4K, at 1.79 Mbps.

Even to the eye, you can’t tell the difference. Concentrate bits only where the human eye looks (commutator-norm), dropped into existing H.264 / AV1 with no decoder change.

VIDEO category → Live A/B demo → 30-day trial →

Patents pending: JP 2026-046898 / 046609 · Paper: IEEE TCSVT v8 submission

• LANGUAGE2026-05-26 ★★★ AS/400 (IBM i) CL verified bit-exact on real hardware ― Original CL run on a real IBM i, output diffed byte-for-byte against SlimeCL’s bash and Rust: 3 / 3 bit-exact. Real hardware caught fixed-length *CHAR and decimal-format gaps — fixed and published.
• LANGUAGE2026-05-24 ★★★ Mainframe DB backend SlimeTree-VSAM — 480× faster than PostgreSQL ― VSAM-compatible native storage in Rust, direct same-host bench against PostgreSQL 16: sequential cursor 480×, random key access 267×. 1B-record overnight batch from PG's 19.5h to 4.4 min, SHA-256 audit chain built in.
• DEVICE2026-05-23 ★★ SlimeTree-RLM (semantic-driven record system) Rust-ported to 272 KB WASM ― An orthogonal layer over an existing system. Single 272 KB WASM, 24× faster than Python, stable under 10,000-unit stress (zero data loss). Runs server-less.
• DEVICE2026-05-23 ★★ SlimeTree-RLM application (AI): hallucination 66% → 22% (-44pt) without modifying any LLM weight ― Layered orthogonally on Qwen3:8b, 100 questions × 3 trials (σ=4%): "plausible lies" reduced to one-third, response 5.8× faster. RLM is not AI-specific; equally usable for non-AI (audit / decision) applications.
• LANGUAGE2026-05-20 PM4 ★★ SlimeNENC engine rewritten in Rust ― Size 1/3.5, harder to reverse-engineer, output bit-exact to the Python baseline.
• LANGUAGE2026-05-20 PM3 ★ Recovering COBOL + Java from mainframe binaries when source is lost ― New SlimeRESCUE-VOS3, 6/6 exact match with the IBM-canonical emulator.
• LANGUAGE2026-05-20 PM2 ★ US federal Medicare pricing → 5-language auto-conversion ― Domain where 1¢ off = federal violation; 675/675 success.
• LANGUAGE2026-05-20 PM SlimeCOBOL reorganised into 5 SKUs ― Today releases SlimeCOBOL-MF / -GnuCOBOL.
• LANGUAGE2026-05-20 1976-born, ~100k-line accounting system: fully converted ― 685/685 identical (100%).
• LANGUAGE2026-05-19 ★ US federal COBOL Validation (FIPS 21-3) cleared ― 10 corpora, 782 files → 5 languages, 3,910/3,910 match.
• META2026-05-18 SIer / Reseller Partner Programme launched ― Online application open.
• LANGUAGE2026-05-18 PSDP 4-way pricing finalised ― See pricing for details.
• LANGUAGE2026-05-14 PSDP released as an independent product line ― Same-language parallelization, separated from the SlimeNENC family.
• LANGUAGE2026-05-11 SlimeASM released ― z/Architecture HLASM + x64 MASM → Java bit-exact.
• VIDEO2026-05-10 NormH.264 v0.4.2 released ― Trial Signature Mode by default; AV1 path moved to NormAV1.
• LANGUAGE2026-05-10 SlimeNatural released ― Software AG Natural / ADABAS → Java + PostgreSQL bit-exact.
• LANGUAGE2026-05-09 SlimeFORTRAN 9-language multi-target completed ―

An AI pipeline that closes in-house ― RLM + local LM + LoRA training, self-contained; external AI is a gated exception. The full RLM picture →

★ DEVICE I/O Highlight ★ AI application

SlimeTree-RLM — a semantic-driven record system, a single 272 KB Rust binary.

Layered orthogonally over an existing system (LLM / decision engine / business rule) to apply semantic-driven constraints and records. A single Rust binary, deterministic, server-less on browser / mobile / embedded. Usable from both AI and non-AI. Recently extended to word-order-invariant on-device routing (in-browser training, zero upload, tamper-evident audit).

SlimeTree-RLM details → Primary materials / white paper → RoleSlot live demo → DEVICE category →

What SlimeTree-RLM does ― the full picture

Semantic-driven routing (D / µ / R)

Mechanically sorts each prompt into D (direct) / µ (damp) / R (residual) to minimize external-LLM calls. 60–80% LLM cost cut, SHA-256 audit chain.

Word-order-invariant routing (RoleSlot)

Case → role-slot preprocessing gives order-independent routing. Trains in-browser, zero upload, tens of KB, no GPU.

Multilingual (case-marking ≡ injection)

Japanese / Korean / Turkish get it natively via case markers; case-less languages like English reach the same order-invariant structure by attribute injection.

Slime-BoySo who puts the case markers in? 💬 tap for the answer▾

Japanese / Korean / Turkish ship the case markers by birth (like ga / o) ― the speaker adds them for free.

Case-less languages like English? ― your local Gemma4 (the on-box teacher) synthesizes the “case particles” after the fact and injects them. Zero upload, all inside the box.

Either way a tiny student memorizes the role slots ― then it runs teacher-free, deterministic.

Deterministic + tamper-evident audit

Same input → byte-identical result. Every step is hash-chained; tampering is detected instantly and runs are replayable.

Hallucination suppression

Without changing any LLM weight, hallucination 66% → 22% (σ=4%, 100Q × 3 trials).

Platform integrations

Wrap RLM over the Meta / X / Google APIs: try a cheap LLM first → RLM re-checks → only the gaps go to a premium LLM. Skips needless expensive calls, cutting LLM cost a further 50–70% (18 public routes).

Clamp (routine & atypical, bit-exact)

Clamp an LLM fluctuation with routine & atypical S-expr rules, bit-exact. Atypical rules are added on the spot in plain language, deploy-free, version-pinned audit.

Conversation → S-expr (Lisp in words)

Say the rule in natural language and the tool translates it to an S-expr (Gemma4). You never write Lisp. Deterministic gates check it (back-translation, number, boundary).

Spreadsheet + templates

I/O anyone can picture. row = request, verdict column = rules. Customize from a template (.slimepkg); golden tests verify before you publish a version.

Integrated platforms ― 18 public routes, cost-tier B complete, R-ratio 28% measured

Meta6 routes publicX6 routes publicGoogle6 routes publicMSPlannedSlackPlannedLINEPlanned

Clamp (spreadsheet / templates)6-pattern showcaseAdd-on saving figuresRoleSlot demo (EN)RoleSlot demo (JA)WASM bit-exact demoIntegrated build (S1–S9)Platform-integration hubMeasured dashboardPrimary materials / white paper

Research: the order-invariant role structure (case → role-slot) improves small-model sample efficiency on real Universal Dependencies data. The gain is specific to order-sensitive small models and vanishes on large LLMs that already internalize it (honest scope). The method is open; the deterministic core is by subscription.

Small, fast, deterministic. Just layer it orthogonally over what you already run.

Sovereignty without ownership ― the thesis & two live demos →

JAVATEL Inc. has delivered surveillance camera systems, video management, and media streaming across Japan since 1997. Headquartered in Nishinomiya, Hyōgo. The Video Intelligence in our company logo is not a recent trend — it is the direction we have walked for more than 29 years.
Partners: Genetec (longest-tenured authorized partner in Japan) / XEOMA AI NVR / Planet Networks (Taiwan) · Why Javatel · Partners · Contact